How to make text to speech sing, beatbox or go crazy
Clips of a robotic voice singing, beatboxing or making helicopter noises have been memes for years. People try the same with modern text to speech and wonder why it just reads the strange letters aloud.
This guide explains where those tricks come from, the documented DECtalk syntax behind them, and what a modern AI voice such as SuperClonica can do instead.
Where the tricks come from
Most singing and noise tricks were made with DECtalk, the classic formant speech synthesizer. The free NASA game Moonbase Alpha speaks its chat with DECtalk, and players famously typed phoneme commands into the chat to make their astronauts sing. Because DECtalk builds sound from rules, you can tell it exactly which sound to make, for how long and at what pitch.
The DECtalk singing syntax
The DECtalk documentation describes a format with a phoneme in square brackets followed by its duration and pitch number in angle brackets: the duration is in milliseconds and the pitch number comes from its tone table, which lists notes and frequencies. For example, the phoneme ow with 400 and 13 in angle brackets holds one vowel sound for 400 milliseconds on one note. Chaining many of these gives melodies; very short, fast sequences give beatbox, motor and “glitch” effects. Depending on the DECtalk build, phoneme input may need to be switched on, so check the manual of the version you use.
Why modern AI voices do not do it
Neural text to speech, including SuperClonica, learns from recordings of real speech and reads text the way a person would. It does not take phoneme, duration or pitch commands, so DECtalk strings come out as letters or are skipped, and it cannot sing on given notes. That is the price of sounding natural.
What you can do with a modern voice
You can still play with the delivery: speed from 0.7 to 2.0 in the free engine, commas and full stops for pauses, exclamation and question marks for melody, and words spelled the way they should sound. For comedy, pick a character voice, voice short lines and cut them together in an audio editor; repeating or reversing short clips there gives many “glitch” effects without any phoneme codes. Hear voices free on their pages with 30 to 100 characters.
Choose the right tool
For retro robot singing, use DECtalk; the DECtalk guide explains where it comes from today. For natural speech in a character or narrator voice, in 31 languages, use SuperClonica: install it with python -m pip install superclonica-cpu, or train your own voice in the studio from a 5 to 10 second recording. Chinese is not supported yet.